James Ulan is PitchBook’s lead emerging tech analyst.
The future of AI is being shaped at Y Combinator‘s Demo Day this week.
The accelerator, led by CEO Garry Tan, features nearly 200 AI startups in its Summer 2024 batch, about 75% of the total cohort. YC has emerged as the most active investor in companies using large language models to perform multistep tasks that require humanlike reasoning. VC investment in LLM agents has rocketed on the back of foundational AI companies investing billions of dollars into powerful new capabilities.
Here are four themes that emerged from our analysis of the cohort and interviews with several founders.
AI robots are coming
Robots can now learn by mimicking human actions thanks to multimodal AI models and vision language models. As a result, we expect dramatic advances in AI-powered robots.
“Being able to train robots by example is a completely different paradigm,” said Jon Miller Schwartz, co-founder of Ultra, a member of the cohort.
Ultra is designing robots to package and label ecommerce products before shipping. Another YC startup, Yondu, is making robots to stock grocery shelves. And Azalea Robotics plans to move luggage at airports using AI robots.
A host of companies are providing data and software to train these machines, including training data startup Sensei and Cerulion, which makes an open-source robot operating system. YC’s new class points to what we believe is a very early trend of AI powering robots that can do tasks they formerly couldn’t learn.
The value proposition to customers is simple: AI robots provide scalable, consistent, low-churn labor that doesn’t require people management. The punchline is that, in certain tasks, AI robots are vying for parity with human labor and will eventually surpass humans as foundational models make huge leaps forward.
AI navigating websites, software like a human
AI is now helping startups browse the internet and navigate software programs. This functionality is the precursor to AI completing entire tasks without human supervision, likely unlocking billions of enterprise value.
Lighthouz AI integrates with legacy procurement software that manufacturers use to track purchases of essential inputs for their industrial processes. The startup uses AI models to pull data from receipts and purchase orders and insert it into the legacy software. Humans review the data, but Lighthouz says it saves procurement staff weeks of time and aims to make fully autonomous procurement agents in the future.
CardLift offers credit card issuers a way to make their cards the default on popular shopping websites. The company uses AI to traverse ecommerce websites and insert the new debit or credit card number into each merchant’s stored cards webpage. The startup currently supports about 60 merchants. Founder Rithwik Pattikonda expects that improved foundational models will grow its merchant pool because they will navigate websites more easily.
Jobs rapidly transforming
YC’s Summer 2024 AI startups are attacking mundane work with greater gusto than prior cohorts. The shift demonstrates that AI will touch every corner of the global economy, potentially rivaling the impact of the internet, cloud computing and mobile phones.
A few examples: Merlin AI is building AI-powered software to run construction businesses; Olive Legal has designed tools to hasten research for medical malpractice attorneys; and Abel Police is automating police paperwork.
Many of these ideas wouldn’t be possible without domain-specific knowledge or the determination to get close with one’s customers. For example, Merlin AI co-located with a construction business for two months. We expect domain-specific startups to comprise a high percentage of future YC classes in the short to medium term.
AI developer tools in full swing
YC’s summer class is rich in tools to help developers better use foundational AI models.
Pipeshift empowers developers to deploy and fine-tune open-source LLMs. Zenbase AI helps users select models and optimize prompts. Coval helps companies simulate and evaluate their AI programs. Wordware provides a sandbox for AI developers to use natural language to build AI apps and quickly turn them into usable APIs.
Collectively, these startups hope to speed AI development times and increase the capabilities of AI agents and software. These improvements will add to the declining inference costs, better latency and new foundational model capabilities we expect from OpenAI, Anthropic, Meta and others.
Featured image of Y Combinator CEO Garry Tan by Harry Murphy/Getty Images
